Docs + Environment pattern: Multi-Turn Training Pattern

[cwing-nvidia]

Background

Users want to train models on multi-turn conversational tasks where the agent handles back-and-forth interactions with a user

Problem

Users need guidance on:

• What "multi-turn" means
• How to structure multi-turn tasks
• Different approaches to simulating user responses
• How to maintain conversation state and handle termination
• How to verify success in multi-turn tasks

Acceptance Criteria

• [ ] Examples and docs that implement the above

Priority

High - common request for agentic training

Related

• Links to "Using LLM to Simulate User Responses" tutorial